Add Qwen3-ASR-1.7B contrib model

[jimburtoft]

Summary

• Adds NxDI implementation of Qwen3-ASR-1.7B (speech-to-text) using decomposed encoder-decoder pipeline
• Traced audio encoder (3 bucket NEFFs: 5s/10s/30s) + NxDI text decoder (Qwen3-VL scatter pattern)
• Full E2E pipeline validated with EXACT MATCH to CPU reference on trn2.3xlarge

Architecture

• Audio Encoder: 24-layer Whisper-like transformer, Conv2D frontend, d_model=1024, output 2048-dim
• Text Decoder: 28-layer Qwen3, hidden_size=2048, GQA 16/8, QK-norm, mRoPE [24,20,20]
• Pipeline: mel spectrogram -> traced encoder -> scatter embeddings -> NxDI autoregressive decode

Performance (trn2.3xlarge, TP=4, LNC=2, SDK 2.29)

Metric	Value
TTFT (5s audio)	27.5ms
TPOT	4.9ms
RTF (30s audio)	0.020x (50x real-time)
Throughput	194 tok/s
WER (LibriSpeech test-clean)	3.06%
Audio throughput (DP=2, TP=2x2)	~46 audio-sec/wall-sec

Validation

• SDK: 2.29 (DLAMI 20260410, NxDI 0.9.17334)
• Instance: trn2.3xlarge (LNC=2)
• Accuracy: EXACT MATCH with CPU greedy decode
• Edge cases: silence, short audio (<1s), long audio (30s), non-speech all handled correctly

Files

contrib/models/Qwen3-ASR-1.7B/
├── README.md                         # Usage, benchmarks, compatibility
├── src/
│   ├── __init__.py
│   ├── modeling_qwen3_asr.py         # NxDI model class (text decoder)
│   └── audio_encoder.py             # Encoder tracing wrapper
└── test/
    ├── integration/
    │   └── test_model.py            # E2E accuracy + performance tests
    └── unit/

Key Implementation Notes

• Text decoder reuses NeuronQwen3VLForCausalLM with scatter_by_index_put for audio embedding injection
• Audio encoder uses a static wrapper (StaticQwen3ASREncoder) that replaces dynamic cu_seqlens with pre-computed block-diagonal attention mask for trace compatibility
• rope_scaling must use "rope_type": "default" (mRoPE applied externally via rotary_position_ids)
• inline_weights_to_neff=True causes accuracy regression for the encoder (cosine 0.724 vs CPU)

Maintainer

Jim Burtoft (@jimburtoft)